mixdog 0.9.3 → 0.9.5

package/README.md

@@ -1,84 +1,158 @@
 # mixdog
 
-Mixdog is a standalone coding-agent CLI/TUI workspace for developers who want
-one terminal experience across multiple LLM providers.
+Standalone coding-agent CLI/TUI for running a multi-provider AI workspace from
+one terminal.
 
-It combines an Ink-based terminal UI, a multi-provider runtime, workflow agents,
-MCP/plugin/skill loading, memory, web search, channel integrations, and focused
-repo tools for reading, editing, testing, and reviewing code.
+Mixdog combines an Ink-based terminal UI, model/provider routing, workflow
+agents, MCP/plugin/skill loading, memory, web search, channel integrations, and
+repo-focused tools for reading, editing, testing, and reviewing code.
 
 ## Highlights
 
-- Multi-provider model routing for Anthropic, OpenAI-compatible APIs, Google,
-  XAI/Grok, OAuth-backed providers, and local endpoints.
-- Full-screen TUI with slash commands, model/workflow pickers, provider setup,
-  usage dashboards, tool cards, statusline integration, and resumable sessions.
-- Agent task delegation through the `agent` tool and `/agent` command.
-- Focused repo tools for `read`, `grep`, `glob`, `list`, `code_graph`,
+- **Multi-provider model picker** for Anthropic, OpenAI, Google/Gemini, xAI/Grok,
+  DeepSeek, OpenCode Go, OAuth-backed providers, OpenAI-compatible APIs, Ollama,
+  and LM Studio/local endpoints.
+- **Live model catalog** from provider `/models` endpoints, enriched with
+  LiteLLM/models.dev metadata for context windows, output limits, pricing, tool
+  support, reasoning, web search, and recency.
+- **Full-screen TUI** with slash commands, provider setup, model/workflow
+  pickers, usage dashboards, statusline integration, resumable sessions, and
+  detailed tool cards.
+- **Workflow delegation** through `/agent` and the `agent` tool, including
+  worker, heavy-worker, reviewer, debugger, maintainer, and explorer roles.
+- **Repo-native tools** for `read`, `grep`, `glob`, `list`, `code_graph`,
   `apply_patch`, `shell`, `cwd`, `explore`, web search, and memory recall.
-- Workflow agents for implementation, review, debugging, maintenance, web
-  research, and heavier model routes.
-- Optional Discord/channel/webhook/schedule integrations for remote or event
-  driven workflows.
+- **Remote/event workflows** via optional Discord channel, webhook, schedule,
+  and channel-management integrations.
 
 ## Requirements
 
 - Node.js >= 22
 
-## Setup
+## Install
 
 ```bash
 npm install
 ```
 
-To install the CLI globally from this checkout:
+To install the `mixdog` command globally from this checkout:
 
 ```bash
 npm install -g .
 mixdog --help
 ```
 
-## Usage
+## Run
 
 ```bash
-# Run the TUI from any project directory
+# Start the TUI in the current project
 mixdog
 
-# Run the basic smoke test
-npm run smoke
+# Start with an explicit route
+mixdog --provider anthropic-oauth --model claude-haiku-4-5-20251001
 
-# Run the core smoke suite
-npm run smoke:all
+# Read-only tool surface
+mixdog --readonly
+
+# Enable remote/channel mode for this session
+mixdog --remote
+
+# Re-run first-run setup
+mixdog --onboarding
+```
+
+Headless role mode is also supported:
+
+```bash
+mixdog worker "fix the failing test"
+mixdog reviewer "review the current diff"
 ```
 
-Inside the TUI, use `/providers` to configure model access, `/model` to switch
-models, `/workflow` to select a workflow, `/agents` to inspect available
-workflow agents, and `/agent` to manage active agent tasks.
+## TUI basics
 
-## Project Layout
+Common slash commands:
+
+```text
+/help              show help
+/providers         configure provider auth and local endpoints
+/model             choose provider/model, effort, and fast mode
+/workflow          choose the active workflow
+/agents            show available workflow agents
+/agent             manage active agent tasks
+/mode              switch tool surface: full | readonly
+/compact           compact older conversation context
+/clear             reset the conversation and screen
+/OutputStyle       show or switch Lead output style
+```
+
+Use `/providers` first if no model is configured, then `/model` to pick the
+route. The model picker warms the provider catalog in the background and keeps
+Claude families such as Opus, Sonnet, Haiku, and Fable separate when filtering
+current Anthropic models.
+
+## Scripts
+
+```bash
+npm run smoke          # basic smoke check
+npm run smoke:all      # core smoke suite
+npm run smoke:tui      # TUI render smoke
+npm run smoke:tools    # tool smoke suite
+npm run build:tui      # build the bundled Ink TUI
+npm run audit:models   # inspect model catalog metadata
+```
+
+Additional diagnostics and benchmarks live under `scripts/`.
+
+## Data and configuration
+
+Mixdog uses `~/.mixdog` as its home root. Runtime data lives in
+`~/.mixdog/data` by default.
+
+```bash
+MIXDOG_HOME=/path/to/home mixdog
+MIXDOG_DATA_DIR=/path/to/data mixdog
+```
+
+Useful environment toggles:
+
+- `MIXDOG_TUI_MOUSE=0` — use terminal-native mouse behavior instead of the TUI
+  mouse/selection layer.
+- `MIXDOG_DISABLE_MODEL_PREFETCH=1` — disable background provider model prefetch.
+- `MIXDOG_PROVIDER_MODEL_WARMUP_DELAY_MS=<ms>` — tune model-catalog warmup delay.
+- `MIXDOG_MODEL_STALE_MONTHS=<months>` — tune catalog staleness filtering.
+
+## Project layout
 
 ```text
 src/
   cli.mjs        # CLI entry point (bin: mixdog)
-  app.mjs        # app wiring
-  repl.mjs       # plain REPL loop
-  runtime/       # provider runtime, tools, memory, channels
-  tui/           # canonical Ink TUI (bundled to dist/)
+  app.mjs        # CLI/TUI/headless mode wiring
+  help.mjs       # command help text
+  runtime/       # providers, tools, memory, channels, session runtime
+  session-runtime/
+                 # model routing, catalog rows, workflow/session helpers
+  tui/           # canonical Ink TUI
   agents/        # workflow agent definitions
+  workflows/     # workflow definitions
   rules/         # Lead and agent instructions
 scripts/
   smoke*.mjs     # smoke checks
+  *test.mjs      # focused node:test checks
   build-tui.mjs  # esbuild bundle for the React TUI
 vendor/
   ink/           # Mixdog Ink renderer
 ```
 
-## Data And Configuration
+## Published package contents
+
+The npm package is limited by `package.json#files` to:
 
-Standalone Mixdog uses `~/.mixdog` as its home root. Runtime data lives in
-`~/.mixdog/data` by default. Set `MIXDOG_HOME` to move the root, or
-`MIXDOG_DATA_DIR` to override only the data directory.
+```text
+README.md
+scripts/
+src/
+vendor/
+```
 
-The TUI uses alt-screen mode and app-owned mouse wheel scrolling / drag
-selection by default. Set `MIXDOG_TUI_MOUSE=0` to opt into terminal-native mouse
-behavior.
+`docs/` is not included in the published package unless `package.json#files` is
+changed.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mixdog",
-  "version": "0.9.3",
+  "version": "0.9.5",
   "private": false,
   "type": "module",
   "description": "Standalone mixdog coding-agent CLI/TUI workspace.",
@@ -26,6 +26,7 @@
   "files": [
     "README.md",
     "scripts/",
+    "!scripts/bench/cache-hit-*.json",
     "src/",
     "vendor/"
   ],
@@ -33,7 +34,7 @@
     "prepack": "npm run build:tui",
     "start": "node src/cli.mjs",
     "smoke": "node scripts/smoke.mjs",
-    "smoke:all": "npm run smoke && npm run smoke:boot && npm run smoke:tools && npm run smoke:output",
+    "smoke:all": "npm run smoke && npm run smoke:boot && npm run smoke:tools && npm run smoke:output && npm run smoke:tui && npm run smoke:freevars",
     "smoke:boot": "node scripts/boot-smoke.mjs",
     "smoke:compact": "node scripts/compact-smoke.mjs",
     "smoke:loop": "node scripts/smoke-loop.mjs",
@@ -41,6 +42,8 @@
     "smoke:loop:report": "node scripts/smoke-loop-report.mjs",
     "smoke:tools": "node scripts/tool-smoke.mjs",
     "smoke:output": "node scripts/output-style-smoke.mjs",
+    "smoke:tui": "node scripts/tui-render-smoke.mjs",
+    "smoke:freevars": "node scripts/freevar-smoke.mjs",
     "test:toolcall": "node --test scripts/toolcall-args-test.mjs",
     "test:providers": "node --test scripts/provider-toolcall-test.mjs",
     "failures": "node scripts/tool-failures.mjs",
@@ -52,6 +55,7 @@
     "bench:corpus": "node scripts/routing-corpus.mjs",
     "bench:run": "node scripts/bench-run.mjs",
     "bench:recall": "node scripts/recall-bench.mjs",
+    "audit:models": "node scripts/model-catalog-audit.mjs",
     "patch:replay": "node scripts/patch-replay.mjs",
     "build:tui": "node scripts/build-tui.mjs"
   },
@@ -70,6 +74,7 @@
     "discord.js": "^14.26.4",
     "ink": "^7.1.0",
     "jsdom": "^26.1.0",
+    "kiwi-nlp": "^0.21.0",
     "marked": "^14.1.4",
     "node-cron": "^4.5.0",
     "openai": "^6.44.0",
@@ -92,6 +97,8 @@
     }
   },
   "devDependencies": {
-    "esbuild": "^0.28.1"
+    "acorn": "^8.17.0",
+    "esbuild": "^0.28.1",
+    "eslint-scope": "^9.1.2"
   }
 }

package/scripts/bench/lead-review-tasks-r3.json

@@ -0,0 +1,20 @@
+[
+  {
+    "id": "lead-review-session-runtime",
+    "agent": "reviewer",
+    "cwd": "C:/Project/mixdog-bench",
+    "prompt": "READ-ONLY review task: you and any delegated agents must NOT modify, create, or delete any files - report findings only. Review the folder src/session-runtime/ for correctness, regression, and security risks. Delegate exploration/reading to your workflow agents where useful and synthesize their findings. Focus on real actionable bugs: broken imports, undefined identifiers, race conditions, wrong fallbacks, data loss paths. Report findings ordered by severity, one line each, with file:line anchors. Skip style nits. End with a one-line verdict."
+  },
+  {
+    "id": "lead-review-channels-lib",
+    "agent": "reviewer",
+    "cwd": "C:/Project/mixdog-bench",
+    "prompt": "READ-ONLY review task: you and any delegated agents must NOT modify, create, or delete any files - report findings only. Review the folder src/runtime/channels/lib/ for correctness, regression, and security risks. Delegate exploration/reading to your workflow agents where useful and synthesize their findings. Focus on real actionable bugs: message loss, double-forwarding, ownership races, watcher leaks, unsafe input handling. Report findings ordered by severity, one line each, with file:line anchors. Skip style nits. End with a one-line verdict."
+  },
+  {
+    "id": "lead-review-compact",
+    "agent": "reviewer",
+    "cwd": "C:/Project/mixdog-bench",
+    "prompt": "READ-ONLY review task: you and any delegated agents must NOT modify, create, or delete any files - report findings only. Review the folder src/runtime/agent/orchestrator/session/compact/ for correctness, regression, and security risks. Delegate exploration/reading to your workflow agents where useful and synthesize their findings. Focus on real actionable bugs: schema validation gaps, data corruption on summarize, lossy round-trips, error swallowing. Report findings ordered by severity, one line each, with file:line anchors. Skip style nits. End with a one-line verdict."
+  }
+]

package/scripts/bench/lead-review-tasks.json

@@ -0,0 +1,20 @@
+[
+  {
+    "id": "lead-review-session-runtime",
+    "agent": "reviewer",
+    "cwd": "C:/Project/mixdog-bench",
+    "prompt": "Review the folder src/session-runtime/ for correctness, regression, and security risks. Delegate exploration/reading to your workflow agents where useful and synthesize their findings. Focus on real actionable bugs: broken imports, undefined identifiers, race conditions, wrong fallbacks, data loss paths. Report findings ordered by severity, one line each, with file:line anchors. Skip style nits. End with a one-line verdict."
+  },
+  {
+    "id": "lead-review-channels-lib",
+    "agent": "reviewer",
+    "cwd": "C:/Project/mixdog-bench",
+    "prompt": "Review the folder src/runtime/channels/lib/ for correctness, regression, and security risks. Delegate exploration/reading to your workflow agents where useful and synthesize their findings. Focus on real actionable bugs: message loss, double-forwarding, ownership races, watcher leaks, unsafe input handling. Report findings ordered by severity, one line each, with file:line anchors. Skip style nits. End with a one-line verdict."
+  },
+  {
+    "id": "lead-review-compact",
+    "agent": "reviewer",
+    "cwd": "C:/Project/mixdog-bench",
+    "prompt": "Review the folder src/runtime/agent/orchestrator/session/compact/ for correctness, regression, and security risks. Delegate exploration/reading to your workflow agents where useful and synthesize their findings. Focus on real actionable bugs: schema validation gaps, data corruption on summarize, lossy round-trips, error swallowing. Report findings ordered by severity, one line each, with file:line anchors. Skip style nits. End with a one-line verdict."
+  }
+]

package/scripts/bench/r4-mixed-tasks.json

@@ -0,0 +1,20 @@
+[
+  {
+    "id": "r4-debug-forwarding-gap",
+    "agent": "reviewer",
+    "cwd": "C:/Project/mixdog-bench",
+    "prompt": "Investigate a real reliability question in this repo: when the Discord channel bridge activates (bridge-state.json written) there can be a window where assistant output is not yet forwarded to the channel - transcript binding, fs.watch on a not-yet-created transcript file, and ownership probes are involved. Trace the full path from bridge activation to first forwarded message in src/runtime/channels/ (index.mjs, lib/output-forwarder.mjs), identify every gap where output produced during the window would be silently dropped or delayed, and state for each gap whether backlog catch-up recovers it. Report findings ordered by impact with file:line anchors. End with a one-line verdict."
+  },
+  {
+    "id": "r4-impl-usage-summary",
+    "agent": "worker",
+    "cwd": "C:/Project/mixdog-bench",
+    "prompt": "Implement a small feature: add a `usage:today` npm script backed by a new scripts/usage-today.mjs that reads ~/.mixdog/data/history/agent-trace.jsonl (respect MIXDOG_DATA_DIR), filters usage rows from local midnight, and prints per-model calls, input/cached/output tokens, and costUsd totals plus a grand total line. Follow the style of scripts/session-bench.mjs arg parsing. Verify with node --check and run it once. Report the diff summary."
+  },
+  {
+    "id": "r4-explain-compact-engine",
+    "agent": "explore",
+    "cwd": "C:/Project/mixdog-bench",
+    "prompt": "Map the compaction subsystem: starting from src/runtime/agent/orchestrator/session/compact/engine.mjs, produce a concise architecture brief - entry points and callers, the decision flow for when compaction fires, how the preserved tail is chosen, where redaction hooks in, and the failure/fallback paths. Anchor every claim with file:line. Keep it under 40 lines."
+  }
+]

package/scripts/bench/r5-orchestrated-task.json

@@ -0,0 +1,7 @@
+[
+  {
+    "id": "r5-orchestrated-audit",
+    "agent": "lead",
+    "prompt": "You are coordinating work in this repo. Task: audit and improve the provider retry story in two independent scopes, then verify. Scope A: analyze src/runtime/agent/orchestrator/providers/retry-classifier.mjs and how openai-oauth-ws.mjs consumes its classifiers - list every classifier bucket, which are actually reachable, and any bucket that no code path can produce. Scope B: analyze the handshake retry loop in openai-oauth-ws.mjs (_acquireWithRetry) and the midstream retry loop (sendViaWebSocket) - document their budgets, backoff, and whether any error class can double-retry across both layers. These scopes are independent: delegate them to separate workers in parallel, then have the results cross-checked, and produce a single merged report with file:line evidence. Do NOT edit any files - analysis only. The final report is the deliverable."
+  }
+]

package/scripts/bench/review-tasks.json

@@ -0,0 +1,20 @@
+[
+  {
+    "id": "review-session-runtime",
+    "agent": "reviewer",
+    "cwd": "C:/Project/mixdog",
+    "prompt": "Review the folder src/session-runtime/ for correctness, regression, and security risks. Focus on real actionable bugs: broken imports, undefined identifiers, race conditions, wrong fallbacks, data loss paths. Report findings ordered by severity, one line each, with file:line anchors. Skip style nits. End with a one-line verdict."
+  },
+  {
+    "id": "review-channels-lib",
+    "agent": "reviewer",
+    "cwd": "C:/Project/mixdog",
+    "prompt": "Review the folder src/runtime/channels/lib/ for correctness, regression, and security risks. Focus on real actionable bugs: message loss, double-forwarding, ownership races, watcher leaks, unsafe input handling. Report findings ordered by severity, one line each, with file:line anchors. Skip style nits. End with a one-line verdict."
+  },
+  {
+    "id": "review-compact",
+    "agent": "reviewer",
+    "cwd": "C:/Project/mixdog",
+    "prompt": "Review the folder src/runtime/agent/orchestrator/session/compact/ for correctness, regression, and security risks. Focus on real actionable bugs: schema validation gaps, data corruption on summarize, lossy round-trips, error swallowing. Report findings ordered by severity, one line each, with file:line anchors. Skip style nits. End with a one-line verdict."
+  }
+]

package/scripts/bench/round-codex.json

@@ -0,0 +1,114 @@
+{
+  "round": "1",
+  "runner": "codex",
+  "opts": {
+    "provider": "openai-oauth",
+    "model": "gpt-5.5",
+    "effort": "high",
+    "fast": false,
+    "env": {}
+  },
+  "tasks": 3,
+  "completed": 3,
+  "completion_rate": 100,
+  "sessions": [
+    "019f24e5-98ee-7cf0-afb0-a1a80ac81702",
+    "019f24e5-98f3-7f12-9638-7a899028c7e3",
+    "019f24e5-98f3-7f22-902c-934f0f7cbfb0"
+  ],
+  "results": [
+    {
+      "id": "review-session-runtime",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 300087,
+      "sessionId": "019f24e5-98ee-7cf0-afb0-a1a80ac81702",
+      "card": {
+        "session": "019f24e5-98ee-7cf0-afb0-a1a80ac81702",
+        "wall_ms": 300087,
+        "uncached_in": 157635,
+        "cached_in": 1071360,
+        "out_tokens": 8581,
+        "cache_ratio": 0.872,
+        "completed": 1
+      },
+      "output": "- High: `src/session-runtime/plugin-mcp.mjs:40` project `.mcp.json` stdio entries can set `cwd` to `..` or an absolute path outside the project, bypassing the stricter project-root guard in `normalizeMcpServerInput` and letting repo-local MCP config run with access scoped outside the opened repo.\n- Medium: `src/session-runtime/workflow.mjs:276` invalid persisted `effort` values throw from read paths like `summarizeWorkflowRoutes`/`agentRouteFromConfig` instead of being ignored, so one bad config entry can break onboarding status or agent listing.\n- Low: `src/session-runtime/provider-models.mjs:42` cached `{id, provider}` fallback metadata is returned before honoring `allowFetch`, so a prior non-fetch lookup prevents a later forced lookup from ever retrieving real model capabilities.\n\nVerdict: no broken imports found, but fix the project MCP cwd escape before trusting repo-local MCP configs."
+    },
+    {
+      "id": "review-channels-lib",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 172813,
+      "sessionId": "019f24e5-98f3-7f12-9638-7a899028c7e3",
+      "card": {
+        "session": "019f24e5-98f3-7f12-9638-7a899028c7e3",
+        "wall_ms": 172813,
+        "uncached_in": 136264,
+        "cached_in": 1237376,
+        "out_tokens": 8502,
+        "cache_ratio": 0.901,
+        "completed": 1
+      },
+      "output": "- High `src/runtime/channels/lib/webhook/deliveries.mjs:255`: delivery rows are appended asynchronously but the in-memory dedup state advances immediately, so crashes, append failures, out-of-order status writes, or compaction can lose/reorder `received/pending/done` rows and allow duplicate webhook dispatch or permanent dedup blockage.\n- High `src/runtime/channels/lib/event-queue.mjs:239`: queued events are moved to processed at `:241` after a fire-and-forget `injectFn`; if parent IPC drops/fails, the claimed event is acknowledged anyway and the message is lost.\n- Medium `src/runtime/channels/lib/output-forwarder.mjs:316`: actual backend send does not re-check `_isOwner()` at send time, so an ownership handoff during an in-flight drain can let the old owner send while the new owner tails the same transcript, causing double-forwarding.\n- Medium `src/runtime/channels/lib/output-forwarder.mjs:690`: `closeWatcher()` clears `sendRetryTimer` but leaves queued failed items in `finalLane/streamLane`, so transcript rebind/watch error during backoff can strand unsent messages until unrelated new activity kicks the drain.\n- Medium `src/runtime/channels/lib/webhook.mjs:530`: `startNgrok()` schedules `waitAndStart` timers that `stop()` at `:585` never cancels, so reload/stop during the delay can spawn a stale ngrok process after shutdown.\n- Low `src/runtime/channels/lib/inbound-routing.mjs:35`: stale dedup marker files are treated as non-duplicates but never refreshed/replaced, so cross-process duplicate inbound messages after TTL can bypass dedup repeatedly.\n\nVerdict: Not clean yet; I found real message-loss, double-forwarding, and lifecycle leak risks worth fixing before shipping."
+    },
+    {
+      "id": "review-compact",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 140945,
+      "sessionId": "019f24e5-98f3-7f22-902c-934f0f7cbfb0",
+      "card": {
+        "session": "019f24e5-98f3-7f22-902c-934f0f7cbfb0",
+        "wall_ms": 140945,
+        "uncached_in": 102061,
+        "cached_in": 416512,
+        "out_tokens": 7120,
+        "cache_ratio": 0.803,
+        "completed": 1
+      },
+      "output": "- High `src/runtime/agent/orchestrator/session/compact/engine.mjs:588` - `recallTailStartIndex()` relies on object identity, so when `fitSingleRecallTurnToCap()` clones/truncates the first tail user message, `indexOf()` fails and the fallback can misplace the tail, duplicating summarized user content and dropping/reordering recent messages.\n- High `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:269` - `redactMessageToolCallSecrets()` returns unchanged unless `m.toolCalls` exists, leaving Anthropic-style `content` / `assistantBlocks` `tool_use` inputs unredacted in preserved tails.\n- High `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:12` - the sensitive-key regex misses common camelCase token keys like `accessToken`, `idToken`, and `sessionToken`, so those secrets survive both prompt summaries and verbatim tail redaction.\n- Medium `src/runtime/agent/orchestrator/session/compact/budget.mjs:79` - `drainSessionCycle1()` swallows `cycle1` errors and returns a “successful” partial drain, so callers can proceed with raw pending rows while diagnostics lose the actual failure.\n- Medium `src/runtime/agent/orchestrator/session/compact/summary-schema.mjs:68` - semantic schema validation only checks top-level `##` anchors, while `truncateSummaryBySections()` can drop required `### Done/In Progress/Blocked` progress structure and still treat the summary as valid.\n\nVerdict: Needs fixes before compaction can be considered secure and loss-safe."
+    }
+  ],
+  "task_complete": true,
+  "task_errors": [],
+  "score_complete": true,
+  "score_errors": [],
+  "group": {
+    "n": 3,
+    "wall_ms": 204615,
+    "uncached_in": 131986.7,
+    "cached_in": 908416,
+    "out_tokens": 8067.7,
+    "cache_ratio": 0.9,
+    "completed": 1
+  },
+  "cards": [
+    {
+      "session": "019f24e5-98ee-7cf0-afb0-a1a80ac81702",
+      "wall_ms": 300087,
+      "uncached_in": 157635,
+      "cached_in": 1071360,
+      "out_tokens": 8581,
+      "cache_ratio": 0.872,
+      "completed": 1
+    },
+    {
+      "session": "019f24e5-98f3-7f12-9638-7a899028c7e3",
+      "wall_ms": 172813,
+      "uncached_in": 136264,
+      "cached_in": 1237376,
+      "out_tokens": 8502,
+      "cache_ratio": 0.901,
+      "completed": 1
+    },
+    {
+      "session": "019f24e5-98f3-7f22-902c-934f0f7cbfb0",
+      "wall_ms": 140945,
+      "uncached_in": 102061,
+      "cached_in": 416512,
+      "out_tokens": 7120,
+      "cache_ratio": 0.803,
+      "completed": 1
+    }
+  ]
+}